보일러 미세먼지 저감을 위한 운전 변수 예측 인공지능 모델의 강화학습 응용

Abstract

Recently, social issues caused by fine dust have threatened the health of people around the world. In particular, sulfur oxides (SOx) and nitrogen oxides (NOx) generated in the process of burning coal are declared as fine dust first-class carcinogens. Already, advanced countries are shutting down coal-fired power plants in stages without expanding them due to damage caused by accelerating global warming and continuous emission of air pollutants. In South Korea, the government plans to suspend new construction of coal power plants and reduce the operation of existing facilities in order to reduce fine dust under the 2020-2040 National Environmental Comprehensive Plan. Recently, major domestic power generation companies have been pushing to reduce fine dust emissions from coal-fired power plants, which are further strengthened than government measures. In addition, it is actively investing in R&D such as development of technologies for improving efficiency of reduction facilities to reduce sulfur oxides (SOx) and nitrogen oxides (NOx), which are major precursors of fine dust. In this study, operating variables are predicted to reduce fine dust in thermal power generation boilers. Locate the value of the operation variable that minimizes SOx and NOx by linking the SOx and NOx prediction models to the Reinforcement learning. Create a NOx and SOx prediction model using the operation information and status information that are the operation variables of the boiler. AutoML (Automated Machine Learning) was used to create the prediction model. AutoML is used as a tool to instead perform experimental actions and trial and error that adjust various parameters that occur in the machine learning model process. In addition, in order to improve the prediction of SOx and NOx, a model that predicts other dependent variables is created and the predicted value is reused as an independent variable. The process of selecting the optimal independent variable was selected by utilizing Variable Importance. It was confirmed that the SOx and NOx model prediction performance is improved when the values of the optimal independent variable and the dependent variable are reused as independent variables. The NOx and SOx prediction models were linked with deep Q-learning, which is one of the reinforcement learning algorithms, to optimize the reduction of fine dust. Within the deep Q-learning environment, SOx and NOx prediction models were used to update status. The action controls the mixture ratio of the boiler combustion carbon model, which is the driving variable. Depending on the action, SOx and NOx values will be returned as rewards. Compensation develops policy in the direction of minimizing SOx and NOx values. For the development of the policy, walked with the ratio of main coal type and auxiliary coal type and the ratio of high calorie bullet and low-calorie coal as constraints. As a result, the amount of main coal species used increased by an average of 3%, and the proportion of lowcalorie coal used increased by an average of about 24%. As shown, the SOx value decreased by 8% on average, and the NOx value also decreased by 8.3% on average. Through the study of reinforcement learning able to consider both economic and environmental aspects in the operation of thermal power boilers at the same time.